Reliable learning-based robot control remains a major challenge for complex, long-horizon tasks, where large volumes of high-quality interaction data are required to achieve the robustness demanded by industrial environments. While low-fidelity (LF), GPU-accelerated simulators enable scalable data generation, their simplified physical models often fail to capture contact-critical dynamics, leading to poor generalization. Conversely, high-fidelity (HF) physics-based simulators provide accurate and predictive modelling of physical interactions but are computationally prohibitive for large-scale policy training. This work introduces a physically anchored multi-fidelity simulation framework that integrates targeted HF simulation with massively parallel LF simulation for efficient and reliable imitation learning. HF simulation is selectively employed to ground and validate critical interaction behaviors, while LF simulation is systematically derived through controlled physical model reduction. Key simulation fidelity axes, including solver configurations, collision and contact representations and simplification, geometry simplification, compliance and actuator models, as well as sensor noise are explicitly parameterized and varied to study their impact on data quality, learning dynamics, and policy robustness. By anchoring LF rollouts to HF-validated reference behaviors, the framework ensures that large-scale dataset generation remains physically meaningful while retaining computational efficiency. The resulting control policies are evaluated through a simulation-driven benchmarking pipeline using both task-level and physics-level metrics, including success rate, force and torque tracking errors, energy consistency, and characteristic failure modes. Performance is quantified in terms of (i) computational cost per simulation episode, (ii) learning efficiency measured by the number of simulated episodes required to reach a target performance threshold, and (iii) final robustness assessed under high-fidelity simulation conditions or on the real robotic system. Overall, the proposed approach demonstrates how structured HF–LF simulation integration enables scalable, physically grounded data generation, significantly reducing simulation cost while producing robust and transferable robot control policies suitable for complex industrial automation scenarios.